57       const float32x2_t vtltr0 = vld1_f32(itl0);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
58       const float32x2_t vblbr0 = vld1_f32(ibl0);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
59       const float32x2_t vtltr1 = vld1_f32(itl1);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
60       const float32x2_t vblbr1 = vld1_f32(ibl1);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
61       const float32x2_t vtltr2 = vld1_f32(itl2);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
62       const float32x2_t vblbr2 = vld1_f32(ibl2);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
63       const float32x2_t vtltr3 = vld1_f32(itl3);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
64       const float32x2_t vblbr3 = vld1_f32(ibl3);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
65       const float32x2_t vtltr4 = vld1_f32(itl4);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
66       const float32x2_t vblbr4 = vld1_f32(ibl4);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
67       const float32x2_t vtltr5 = vld1_f32(itl5);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
68       const float32x2_t vblbr5 = vld1_f32(ibl5);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
69       const float32x2_t vtltr6 = vld1_f32(itl6);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
70       const float32x2_t vblbr6 = vld1_f32(ibl6);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
71       const float32x2_t vtltr7 = vld1_f32(itl7);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
72       const float32x2_t vblbr7 = vld1_f32(ibl7);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
137       const float32x2_t vtltr0 = vld1_f32(itl0);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
138       const float32x2_t vblbr0 = vld1_f32(ibl0);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
139       const float32x2_t vtltr1 = vld1_f32(itl1);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
140       const float32x2_t vblbr1 = vld1_f32(ibl1);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
141       const float32x2_t vtltr2 = vld1_f32(itl2);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
142       const float32x2_t vblbr2 = vld1_f32(ibl2);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
143       const float32x2_t vtltr3 = vld1_f32(itl3);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
144       const float32x2_t vblbr3 = vld1_f32(ibl3);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
180         const float32x2_t valphah = vw.val[0];  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
181         const float32x2_t valphav = vw.val[1];  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
189         const float32x2_t vtltr0 = vld1_f32(itl0);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
190         const float32x2_t vblbr0 = vld1_f32(ibl0);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
191         const float32x2_t vtltr1 = vld1_f32(itl1);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
192         const float32x2_t vblbr1 = vld1_f32(ibl1);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
194         const float32x2_t vldrd0 = vsub_f32(vblbr0, vtltr0);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
195         const float32x2_t vldrd1 = vsub_f32(vblbr1, vtltr1);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
198         const float32x2_t vld = vld_t.val[0];  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
199         const float32x2_t vrd = vld_t.val[1];  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
202         const float32x2_t vtl = vtl_t.val[0];  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
203         const float32x2_t vtr = vtl_t.val[1];  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
205         const float32x2_t vl = vfma_f32(vtl, vld, valphav);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
206         const float32x2_t vr = vfma_f32(vtr, vrd, valphav);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
208         const float32x2_t vd = vsub_f32(vr, vl);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
209         const float32x2_t vo = vfma_f32(vl, vd, valphah);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
229         const float32x2_t valphav = vld1_dup_f32(w + 1);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
236         const float32x2_t vtltr = vld1_f32(itl);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
237         const float32x2_t vblbr = vld1_f32(ibl);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
242         const float32x2_t vldrd = vsub_f32(vblbr, vtltr);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
243         const float32x2_t vlr = vfma_f32(vtltr, vldrd, valphav);  in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()